Design of a Predictive Modeling System for MRI Brain Image Classification

The primary organ of the human body is the brain, and it is still unclear what causes brain cancer. Consequently, an early diagnosis is necessary to lower the fatality rate from brain cancer. One diagnostic method utilized in the medical profession that does not involve segmentation is image classification. It is regarded to be one of the essential tasks for computer vision. It is accompanied by supervised learning, which is a method of learning in which unstructured data, such as images, are categorized into predetermined classes (also known as labels), which are accessible throughout the training process. Image classification methods have traditionally been based on humanly built features, which need a high degree of domain expertise while displaying poor cross-domain adaptability. More recently, automated methods have begun to replace these manual methods. MRI brain images have several challenges which makes the conventional machine learning techniques less effective. They are summarized as below:

•High Dimensionality: MRI scans are high-resolution resulting in a substantial quantity of data for each sample.

•Complex features: The abnormalities are often subtle and may vary from patient to patient.

•Intra-class Variability: Different patients with the same condition can have variable degrees of abnormalities in MRIs, which makes classification more difficult.

Inter-class Similarity: Due to the similar visual appearances of healthy brin tissue and abnormal tissue structures, the classification task is very challenging. In recent years, Deep Learning (DL) has been used to solve complex issues that are found in the real world. The abovementioned challenges are addressed by the DL and are as follows:

•Automatic Feature Extraction: Feature extraction by traditional machine learning algorithms is subjective and incomplete, particularly when dealing with complicated data such as MRI images. The ability of DL architecture to automatically learn hierarchical features directly from raw images is a significant advantage.

•Handling High-Dimensional Data: DL models can process three-dimensional data efficiently, using CNNs or autoencoders, to learn spatial relationships within volumetric data.

•Generalizing Across Intra-Class Variability: DL models have the ability at generalizing across big datasets, learning robust features that can account for intra-class variability.

•Discriminating between Inter-class Similarity: DL models learn highly discriminative features, which means that they can discover small distinctions across similar classes of raw images.

DL is a machine learning method that processes information non-linearly through hierarchical structures [1]. Its goal is to identify patterns and extract features from data. DL utilizes Artificial Neural Networks (ANN) [2] to mimic human brain functions and learn complex characteristics from input data. The proposed system applies DL for both binary (normal/abnormal) and multiclass (normal/low-grade/high-grade) classifications in brain image analysis, using non-invasive Magnetic Resonance Imaging (MRI) [3].

The main objective is to create a classification system for brain cancer using MRI scans with high accuracy and capable of early detection. The rest of the paper is as follows. Section 2 reviews the related work of MRI brain image classification models and methods which utilize DL. The development of a classification technique to classify MRI brain images is covered in Section 3. In section 4, it is demonstrated that the proposed PMS-BCC system is the most accurate system for classifying brain cancer. The key components of the suggested system are outlined in the last section.

This section discusses the proposed PMS-BCC architecture for MRI brain image analysis. Figure 1 presents the workflow of the proposed system that takes an MRI brain image as input. The proposed PMS-BCC architecture uses a DL concept which employs hierarchical structures and several steps of nonlinear information processing to classify patterns and learn features. To localize the image features, stacked convolution and pooling layers with random skip connections are employed. The convolution layer extracts localized features whereas pooling layers reduces feature dimension in the feature extraction module. In the next module the best subset of features is selected by ALO and finally neural network with GLT is employed for the classification.

1.png

Figure 1. Proposed ISM-DR classification system

3.1 Feature extraction module

Figure 2 shows the feature extraction module of the proposed PMS-BCC architecture. It shows the structural details of the arrangement of convolution and pooling layers along with other modules of the proposed system.

2.png

Figure 2. Feature extraction module of the proposed PMS-BCC architecture

The proposed PMS-BCC architecture combines the arrangement of convolution and pooling layers to localize image features from VGG16 architecture [14] with residual mapping from ResNet design [15]. It organizes its layers into five distinct blocks. Each block has two or three 3x3 convolutional layers that are stacked one after the other, followed by a 2×2 max-pooling layer with a stride of 2. This arrangement can gradually capture and understand the complex patterns in the brain images while compressing the spatial dimensions. As the blocks get deeper and deeper, this hierarchical feature extraction is important to the network's success in MRI brain image classification. However, adding more layers (i.e., making structures deeper) results in a vanishing gradient problem, making it harder and harder to train the architecture and causing its accuracy to saturate and then degrade. When the deep networks are trained using gradient-based learning approaches, the vanishing gradient issue arises.

To address the problem of vanishing gradient, Microsoft Research in 2015 [15] introduced the concept of skip connection between the layers. Using the skip connection, training may be skipped for a few levels and linked straight to the output. Hence, the network fits the residual mapping rather than having the layers learn the underlying mapping. If any layer causes the model's performance to suffer in any way, the skip connection will allow regularization to work around the problem. This is one of the advantages of incorporating this kind of connection. Because of this, networks may be trained without the problems that are often produced by vanishing gradients.

3.2 Feature selection module

In this module, ALO is employed which is an artificial intelligence-based feature selection algorithm. The hunting behavior of antlions, which are insects that are often found in sandy areas, was the idea for the ALO method [16]. These insects are famous for their clever approach to seeking their prey. Antlions dig conical holes in sandy areas and wait at the bottom of the holes for unsuspecting insects to fall in. Simulating the hunting behavior of antlions is employed in this work to obtain the best feature subset for brain image classification. The random walk of ants is defined by:

$\begin{gathered}X(t)=\left[0, \text { cumsum }\left(2 \gamma\left(t_1-1\right)\right), \text { cumsum }\left(2 \gamma\left(t_2-\right.\right.\right.  1)), \ldots, \left.\text { cumsum }\left(2 \gamma\left(t_T-1\right)\right)\right]\end{gathered}$   (1）

where, $t$ represents the iteration number, $T$ is the maximum number of iterations, and $\gamma$ is defined by:

$\gamma(t)=\left\{\begin{array}{ll}1 & \text { if rand }>0.5 \\ 0 & \text { if rand } \leq 0.5\end{array}\right\}$     (2)

In Eq. (2), rand is a random number uniformly distributed in (0,1). The position of the ants needs to be regulated so that the ants do not violate the boundaries and move in the search space randomly. The regulation is defined by:

$x_i^t=\frac{\left(x_i^t-a_i\right) \times\left(u b_i^t-l b_i^t\right)}{\left(b_i^t-a_i\right)}+l b_i^t$    (3)

where, $u b_i^t$ and $l b_i^t$ are the maximum and minimum value of certain ant in the $t^{\text {th }}$ generation. $b_i$ and $a_i$ are the maximum and minimum value of the $i^{\text {th }}$ variable of all ants. The calculation for the ants that fall into the trap are given by:

$l b_i^t=$ Antlion $_j^t+l b^t$   (4)

$u b_i^t=$ Antlion $_j^t+u b^t$ （5）

To make the ants move towards the antlions, the ant’s random walk range is lowered adaptively by:

$l b_i^t=\frac{l b^t}{1+10^{v \times \frac{t}{T}}}$           （6）

$u b_i^t=\frac{u b^t}{1+10^{v \times \frac{t}{T}}}$ （7）

where, v depends on the iterations and given by:

$t> \begin{cases}0.1 T & v=2 \\ 0.5 T & v=2 \\ 0.75 T & v=2 \\ 0.9 T & v=2 \\ 0.95 T & v=2\end{cases}$

The position of the ants is updated using the following formula:

$A n t_i^t=\frac{R_l^t+R_E^t}{2}$   (8)

where, $R_l^t$ and $R_E^t$ are position of the ants by route wheel selection and random position around the elite antlion respectively at the $t^{\text {th }}$ iteration. The position of the ants is updated and replace an antlion with its corresponding ant it if satisfies Eq. (9).

Antlion$_j^t=$ Ant$_i^t$ if $f({Ant}_i^t)>f(Antlion_j^t)$   (9)

The new value of the ant's objective function is calculated and compared to the value of the elite. If the ant's fitness value is lower than that of the antlions, the antlions are thought to have caught the ant, and the position of the antlions is updated. The objective function for the proposed PMS-BCC architecture is defined by:

$f(X)=w \times C E(X)+(1-w) \times \frac{\left|X_n\right|}{N}$   (10)

where, $f(X)$ is the fitness function, $N$ is the total number of features (population size), $C E\left(X_i\right)$ is the classification error rate using the subset of X, $X_n$ is the number of selected features and $w=0.8$ is a constant controlling the classification performance to the number of features used. The hyperparameter tuning by ALO is outlined below.

3.png

Figure 3. Structure of neural network

The proposed system randomizes the weight and bias parameters before learning starts using, He initialization. For ‘n’ input values, this approach draws values from a distribution with 0 (zero) mean and variance. It maintains the variance of activations across the network which is a requirement for the rectified linear activation function at hidden layers. The training process is repeated until the needed values and the desired output are attained for both parameters. The two parameters have different effects on the input data. Bias is the difference between the intended and expected values and is responsible for the difference between the network's actual and intended output. A low bias value means that the network generates fewer output form assumptions, while a high bias value means that the network generates more output form assumptions. High bias models are unable to accurately represent the salient characteristics of dataset samples and are ineffective when applied to fresh data. Weights of the networks may be thought of as the degree of neural connectivity. The weight of an input variable determines how much of an impact it has on the output. While a high weight number will have a bigger influence on the output, a low weight value will have minimal impact on the input. A single neuron in the neural network is defined as:

$\hat{y}=f\left(\sum_{k=1}^M x_i w_i+b\right)$   (11)

where, f represents the activation function, M represents the input samples, w and x denoted the weights and inputs to the neuron with bias (b). In this work, rectified linear unit is employed as f. It might be challenging to train neural networks that are extremely deep and contain many layers.

When the number of hidden layers is raised, the amount of inaccurate information that is sent back to earlier levels is dramatically reduced. This shows that although hidden layers near the input layer have scarcely had any changes, layers near the output layer have undergone the necessary updates. This problem, known as the vanishing gradient, made it difficult to train exceedingly deep neural networks. An important turning point in the revival of neural networks was the development of the greedy layer-wise pre-training strategy, which is more often referred to as simply pre-training. This approach is originally accountable for enabling the building of deeper neural network models

Pre-training entails modifying a model while maintaining the weights, adding a new hidden layer, and letting it learn from the prior layer. To overcome the more challenging problem of training a deep network, the model is trained layer by layer. The whole training procedure is divided into a series of layer-wise training using a greedy shortcut to get locally optimal answers [17]. This quick cut results in a workable, worldwide solution. A shallow network may be trained more rapidly than a deep network using pre-training, which employs a layer-wise training method. Shallow networks are thought to be simpler to train than deep networks.

Pre-training may be approached in two main ways: supervised and unsupervised. A model trained via supervised learning is then put through supervised pre-training, which involves progressively introducing hidden layers. The greedy layer-wise method is used in unsupervised pre-training to build an unsupervised auto-encoder model. A supervised output layer is added after unsupervised pre-training. Unsupervised pre-training may be the optimal option if more unlabeled cases are used to initialize the model. This is so that unsupervised pre-training may take place without any human supervision. It is customary to alter the network's weights after the preceding layer is included, even though the weights in the layers underneath the final one are always maintained constant. During training, Mean Square Error (MSE) is used as an objective function which is defined by:

$LossFunction(M S E)=\frac{1}{n} \sum_{i=1}^n\left(y_i-\hat{y}_i\right)^2$   (12)

where, $\hat{y}$ is the network's output and $y$ is the ground truth data. Table 1 shows the parameter setting for the proposed PMSBCC architecture.

Table 1. Parameter setting for the proposed PMS-BCC architecture

At first, the learning rate is set to 0.1. It is reduced by a factor of 10 (0.01, 0.001, 0.0001) when the validation accuracy stopped improving. It is observed that the learning rate is reduced 4 times during training of PMS-BCC architecture and the learning of deep features for the classification of brain images stopped at 245 epochs. There are many activation functions employed in DL architecture for computer vision applications. The commonly used hidden layer’s activation functions are rectified linear, leaky and parametric rectified linear, tanh, exponentially linear unit. The rectified linear function allows only the positive values and rejects all negative values to avoid vanishing gradient problem. Among different functions, rectified linear is widely used due to its low complexity and simplicity. If any dying problem (neurons outputting zero for all inputs) due to rectified linear, the modified rectified linear functions can be used. The proposed PMS-BCC architecture uses rectified linear at hidden layer and SoftMax layer at the output layer as the system is a multiclass classification system.

As the REMBRANDT database has an enormous collection of images, the selection of images is crucial for the system’s performance. The proposed PMS-BCC architecture uses 200 images which was carefully selected for MRI brain image classification in study [21]. The high quality and clarity of the selected images are well annotated and have traces of tumors. Axial, coronal, and sagittal views are commonly employed in medical imaging to indicate the position and direction of structures within the body. The axial view provides a detailed cross-sectional image of the brain compared to others. The coronal and sagittal views provide a clear view of the frontal structure and structures in the left-right orientation respectively. The main limitation is that the proposed PMS-BCC system uses only axial views of brain in the REMBRANDT database.

As the proposed PMS-BCC system is trained on REMBRANDT database, it might not perform well on other datasets due to differences in image acquisition, scanning view, population size, diversification of images per classes. Though the proposed system uses data augmentation to avoid overfitting issuing, regularization methods, cross-validation, and the utilization of larger, and more varied datasets can effectively mitigate the problem of overfitting. Regular assessment on an independent validation set during the training process can also aid in identifying and addressing overfitting at an early stage.